@treeseed/dsh-chapters 0.2.5 → 0.2.6

package/docs/autorun-guide.md

@@ -0,0 +1,248 @@
+# The autorun guide — goal-directed sessions that keep themselves moving
+
+> Status: draft for the maintainer's process review (#186). Promotion to the documentation site
+> (a numbered section beside §28) happens after that review. Every command verb, config key, and
+> default named here was read out of the shipped source at 0.2.3, not remembered.
+
+## 1. What this system is for
+
+A coding agent stops when its turn ends. Everything in this guide exists to close the gap between
+"the work needs forty turns" and "a human is pressing Enter forty times" — without handing the
+machine unchecked authority over the objective:
+
+- **The goal** is the standing directive: what done means, and the stopping rules around it. The
+  host's goal driver re-prompts the model at every turn boundary until the goal is complete,
+  blocked, cleared, or its round budget is spent.
+- **The Suggested Next Step bar** is a second model — the *prompt model*, whose whole job is to
+  write the ONE next instruction — consulting the session's live tail and its archived-chapter
+  summaries after each completed turn.
+- **Autorun** is the clutch between them: armed, each suggestion runs itself as a human-authored
+  turn, so the session keeps working without you. With a goal armed, arming autorun *displaces*
+  the goal driver's own continuation — the goal remains the objective, but the suggestion loop
+  becomes the engine. Without a goal, autorun simply keeps iterating suggestions.
+- **Chapters** make the loop sustainable: when context pressure arrives, the conversation is
+  archived verbatim into chapters and the work continues in a fresh session seeded with nothing
+  but a table of contents. A 32K-context model can run a hundred-turn objective because its
+  context is renewable.
+- **Enrichment and rule distillation** make the loop *learn*: idle-time model passes annotate the
+  archived chapters (title, summary, topics — so retrieval works), and mine them for standing
+  orders (rules with an honest warrant: `stated` only when a human literally said it, `inferred`
+  when the model judged it useful). Rules then assemble into every future task's context.
+
+The two-by-two that orients everything else:
+
+| | no autorun | autorun armed |
+|---|---|---|
+| **no goal** | ordinary chat; the bar offers suggestions, you press Run (or ignore them) | the suggestion loop iterates on its own, up to its round budget |
+| **goal armed** | the host's goal driver re-prompts each turn boundary (rounds cascade until complete/blocked/cleared/budget) | the suggestion loop drives instead — the goal driver's continuation is displaced, the goal stays the objective |
+
+## 2. How it works, turn by turn
+
+1. A turn completes. If a prompt model is set, it is asked for the next step: one imperative
+   instruction, at most `promptMaxChars` (default **240** today) characters, composed from a
+   `promptContextChars` (default **4000**) budget covering the session's chapter summaries, the
+   live conversation tail, and the active goal. An over-cap answer is *refused, never truncated* —
+   a suggestion you can't see whole is a suggestion you can't judge.    As of 0.2.4 (#188, shipped): the suggestion cap defaults to **4096** characters, and both caps
+   belong to the model — `/chapters-prompt model <provider/model> [maxChars] [contextChars]` stores
+   explicit numbers beside the model, and with no explicit numbers the context budget **derives from
+   the model's real window** (`window × promptContextRatio`, default 0.5, at 4 chars/token — a 32K
+   model and a 1M model get budgets 30× apart from one rule). The input itself became the session's
+   memory: the goal whole, the active plan, the core rules, every chapter on the ancestor path as
+   title + detailed summary + every plot line, and the last turn whole — with **zero truncation** as
+   an invariant: over budget, the oldest chapter drops (announced) until the history fits, and a
+   protected set that cannot fit refuses with the numbers instead of sending a starved input.
+2. The offer is classified. Plain text = a **worker-step** (autorun may fire it). A `GOAL:`
+   prefix = a **goal-edit**: the model is proposing to re-cut the objective itself. Autorun must
+   never act on a goal-edit — the host reserves objective changes for direct human turns — so it
+   waits in the bar for you to read and Run yourself.
+3. If autorun is armed and the offer is a worker-step, the loop fires **at most one turn per
+   offer**, submitted through the same path as the bar's Run button: a human-authored user turn.
+   Fired rounds count against `promptMaxRounds` (default **24** per session); offers nobody armed
+   never spend the budget.
+4. Arming autorun while a goal is active calls the host's goal *disarm* seam: the goal driver
+   stops re-injecting its rounds while the goal itself stays durable and unchanged (phase,
+   objective, revision all intact). You can see it: the goal bar's `Pause goal` becomes
+   `Resume goal`. One engine owns each turn boundary — never two.
+5. Context pressure arrives eventually; that is the design, not a failure. The host's relief
+   compacts the surface, chapters archive the span verbatim (oversized tool results defer to
+   content-addressed artifacts), and a continuation opens with the TOC notice as its only seed.
+   The goal rides the session tree into the continuation; **autorun does not** — armed state is
+   process-local per session and nothing arrives armed after a restart or in a child. Re-arm
+   deliberately; a loop that starts turns must never resume itself.
+6. While the session is idle, enrichment drains its queue in batches (`enrichmentIdleMs`,
+   `enrichmentBatchCap`): one structured call per archived chapter asks for
+   `{title, summary, topics, rules[]}` as validated JSON — retry once, and on second failure the
+   deterministic values stand (a slow or failing model never blocks or blanks a chapter). The
+   chapter's body is never touched; annotations land in frontmatter with a `generated:`
+   provenance chain.
+7. Mining rides the same call (or a dedicated one when `rules.model` is set — one extra bounded
+   call per chapter, which is exactly what you pay a better model for). Each candidate rule
+   carries a `quote`; the quote is checked **verbatim against human bytes only** — a quote found
+   in assistant prose is the model restating itself, never evidence you said it. That decides the
+   warrant: `basis: stated` (your words are behind it) or `basis: inferred` (the model judged it
+   useful and nobody said it). With `rulesAutoActivate` on (the default), mined rules govern the
+   session tree on arrival and `revoke` is the undo; with it off, they are held `reviewOnly`
+   until this machine approves them.
+8. Rules close the loop: at the start of every task the model can assemble the standing orders in
+   force (`chapters_rules`), and rebuilt heads render the core set — so what the loop learned
+   yesterday shapes what the loop suggests tomorrow.
+
+## 3. Before you start
+
+1. **Install/update** the plugin in your harness profile and restart the harness:
+   `npm i @treeseed/dsh-chapters@latest` (site §23 covers every path).
+2. **Optionally link a knowledge pool** so chapters and project rules are shared across machines:
+   `/chapters-link <path-or-url>`. Everything in this guide works without it; with it, project
+   rules and the archive travel.
+3. **Pick your models** — this is what `/chapters-models` exists for:
+
+   ```
+   /chapters-models                      # providers + advertised models as paste-ready ids
+   /chapters-models <provider> <model>   # one route: context window, reasoning efforts
+   ```
+
+   The footer shows what each of the three settings currently resolves to. Three knobs, three
+   different jobs:
+
+   | setting | command | job | default |
+   |---|---|---|---|
+   | annotation | `/chapters-enrich model <provider/model｜clear>` | summarize archived chapters for retrieval | unset = rides the conversation route |
+   | mining | `/chapters-rule model <provider/model｜clear>` | distill standing orders — judgment-heavy | unset = rides the annotation call |
+   | prompt | `/chapters-prompt model <provider/model>` | **defines the next work**; setting it IS enabling the bar | unset = the feature does not exist |
+
+   The bar's visibility rule is exactly this setting: with no prompt model, the feature does not
+   exist anywhere; the moment one is set, the bar appears docked above the composer (its own dock
+   slot, order 30) and lives there per session — `Analyzing…` after each completed turn, then
+   `Recommended Next Step`. `/chapters-prompt off` hides it for the session (`on` brings it back),
+   and an idle bar stays visible on purpose: "live, nothing to say yet" must not look like
+   "disabled".
+
+   Tuning intuition: annotation is volume work — a cheap, fast model is right. Mining reads
+   between the lines for instructions worth keeping — this is where a stronger model earns its
+   extra call. The prompt model writes the instruction that *becomes the next turn*; under
+   autorun it is effectively the loop's director, so give it the best model you can afford.
+4. **Decide your review posture.** `rulesAutoActivate: true` (default) means mined rules govern
+   on arrival — fast, and `revoke` undoes any mistake per machine. Setting it `false` holds every
+   mined rule inert until `/chapters-rule approve <id>` — right for shared pools where a
+   mis-generalization would bind teammates. Config keys live on the preset row in your profile's
+   `cordis.patch.yml` (site §16).
+
+## 4. The arc: from a need to a completed goal
+
+**Phase 1 — Identify the work.** Either bring a decision you made outside the session ("we're
+migrating the auth layer; here's the plan") or think it through with the model by ordinary
+conversation. Either is fine; what matters is that the objective ends up stated in words a
+stranger could judge completion by. When the model needs a decision from you it should ask with
+the question tool — which suspends until you answer — rather than burying a question in prose
+and proceeding on a guess.
+
+**Phase 2 — Arm the goal.** Say it plainly: *"Create a goal: <objective>."* The model calls
+`create_goal`; the host's goal bar appears with `Pause goal`. Two numbers matter:
+`max_goal_rounds` caps the automatic continuation rounds (default **256** — a round is itself a
+turn, so rounds cascade until something stops them); ask for a smaller budget when the objective
+is bounded. The goal can be edited, paused, resumed, completed, or marked blocked through the
+model (`update_goal`) — blocked requires the same concrete condition to persist for at least
+three consecutive rounds, so it can't be used to give up early.
+
+**Phase 3 — Hand over the wheel.** With the bar's model set (step 3 above), arm the loop:
+
+```
+/chapters-prompt autorun on
+```
+
+The reply tells you which engine you got: with a goal armed it reports the displacement ("the
+goal remains the objective; its own auto-continuation is disarmed, so the suggestion loop
+drives"), and the goal bar flips to `Resume goal`. Autorun needs a prompt model — arming without
+one is refused, loudly.
+
+**Phase 4 — The loop.** Now watch it work. The bar is your primary surface here — everything
+per-turn happens on it, and the commands you typed in Phase 3 were configuration, not the
+workflow. After each completed turn the bar shows *Analyzing…* then *Recommended Next Step*;
+expand it to read the offer in full; **Run** submits it as your own turn; **Autorun** arms the
+loop so worker-step offers run themselves; **Off** hides the bar for the session. Armed autorun
+fires worker-steps as human-authored turns; pressure points become checkpoints and continuations;
+idle time annotates and mines. Your role shifts from author to editor — which brings us to the
+situations.
+
+**Phase 5 — Steering (everything in between).** See the table in §5. The one rule of thumb:
+**typing anything takes the wheel.** A human turn disarms autorun immediately (the offer stays
+visible, and the reply says why), so interrupting is always safe and never loses the suggestion.
+
+**Phase 6 — Completion.** The model marks the goal complete only when the objective is actually
+achieved, with evidence gathered before the claim; you'll get a closing report. Or it marks
+blocked with the concrete persisting condition — and **blocked disarms autorun exactly like
+complete does** (0.2.4, #188: both planes stop together; a loop firing offers at a dead goal
+outlived its purpose). You can end it yourself any time (`/goal clear`, or Pause on the goal
+bar). While a goal is active or paused, the round budget **stands down** — the pursuit is bounded
+by the goal's lifecycle, never cut off by `promptMaxRounds`. With no goal (or after one is
+cleared), an armed loop keeps iterating suggestions within the budget (default 24): open-ended
+iteration is a feature, but know it's there — the last thing you do is usually `autorun off`.
+
+**Phase 7 — After.** The work outlives the session: chapters are searchable (`chapters_search`
+for the model, the pool for your team if linked), artifacts remain addressable by hash, and mined
+rules persist — session-tier ones stay machine-local and bind that session tree; promote a good
+one to the shared tier with `/chapters-rule add <category> "<text>"` (a new project-tier rule)
+and `approve` it per machine.
+
+## 5. Situations and possible actions
+
+| Situation | What's happening | Action |
+|---|---|---|
+| The suggestion is wrong | the prompt model misread the state | Just type your own instruction — a human turn disarms autorun and the offer stays visible. Or Run a different pace: `autorun off`, steer manually |
+| The suggestion starts with `GOAL:` | the director proposes re-cutting the objective — autorun will never fire these | Read it. If right, press Run yourself (objective changes need a human turn — that's the host's rule, not a limitation of the loop). If wrong, ignore it; the next turn gets a fresh offer |
+| The loop is circling | offers repeat variations; no state changes | `autorun off`, then either re-cut the goal (ask the model to `update_goal` edit) or give one concrete instruction that breaks the cycle |
+| The objective is stale | the work changed shape | Ask the model to edit the goal (a revision bump invalidates outstanding offers — last cycle's directive can't fire against a re-cut objective), or `/goal clear` and create a new one |
+| It won't stop | rounds cascade by design | Three brakes, finest first: `autorun off` (loop), `Pause goal` / ask to pause (driver), round budgets (`promptMaxRounds` 24, `max_goal_rounds` at create) |
+| A mined rule is wrong or over-general | inference happens; that's why the warrant exists | `/chapters-rule list --scope session` → `view <id>` (full text + basis) → `revoke <id>`. Machine-local, one command, file bytes untouched |
+| A mined rule is good and general | worth binding beyond this tree | `/chapters-rule add <category> "<text>"` creates a project-tier proposal; `approve <id>` makes it core on your machine; the pool carries it to teammates |
+| You keep repeating yourself | say it once, durably | Start a line with `ADD RULE: <text>` in your own message — deterministic capture, zero model tokens, active for your session tree, exactly your words |
+| Mined rules worry you | shared pool, high stakes | `rulesAutoActivate: false` in config → REVIEW posture: every mined rule waits inert for `approve`; `/chapters-rule active` shows the posture and what a rebuilt head will render |
+| Annotations are shallow | annotation model too weak, or wrong for the corpus | `/chapters-enrich model <better/id>`; re-run with `/chapters-enrich run` |
+| Mining misses the point | mining rides a cheap annotation call | `/chapters-rule model <stronger/id>` — one extra bounded call per chapter, the intended lever |
+| Enrichment backlog | idle batches haven't drained | `/chapters-enrich report` (status + pending count), `/chapters-enrich run` (drain now) |
+| Session feels long/slow | pressure is normal; relief is automatic | Nothing required — checkpoints and continuations are the design. Optionally `/chapters-save` (archive without a child) or the message doors: **Continue with chapters** (same line of work, fresh context) / **Fork with chapters** (branch an experiment; the parent stays untouched) |
+| The bar is in the way | you want a quiet session | `/chapters-prompt off` hides it; `on` brings it back. Off is per session |
+| Something looks broken | sync, store, or diagnostics | `/chapters-status` (prints the whole diagnostics block in every state); `/chapters-repair` for archive damage (supersedes, never rewrites); `/chapters-rule conflicts` for contradictory standing orders (report-only — it never edits a rule) |
+| Restarted the harness | armed state is process-local by design | The goal persists (durable); autorun does not (nothing arrives armed — a loop that starts turns must never resume itself). Re-arm deliberately |
+
+## 6. Command summary
+
+```
+Discovery          /chapters-models [<provider> [<model>]]
+Suggestions        /chapters-prompt model <provider/model> | on | off | state | autorun on|off
+                   /chapters-prompt suggest   (compose a fresh offer now — the pull the bar rides)
+                   /chapters-prompt run       (submit the offer on screen — the bar's Run button as a command;
+Enrichment         /chapters-enrich run | model <provider/model|clear> | report
+Rules              /chapters-rule add <category> <text> | list [--all|--proposed|--category <c>|--scope project|session]
+                   /chapters-rule view <id> | active | approve <id> | revoke <id> | pin <id> | unpin <id>
+                   /chapters-rule model [<provider/model|clear>] | conflicts
+Archive & health   /chapters-status | /chapters-save | /chapters-repair | /chapters-link
+                   both are session-scoped: "the offer on screen" is a per-session fact)
+Doors (UI)         Continue with chapters · Fork with chapters · the Suggested Next Step bar (expand, Run, Autorun, Off)
+Goals (host)       the goal bar (Pause/Resume) · /goal · the model's create_goal / get_goal / update_goal tools
+Marker (chat)      ADD RULE: <text>   — at the start of a line in your own message
+```
+
+Config keys you're most likely to touch (preset row in `cordis.patch.yml`, site §16):
+`promptMaxRounds` (24), `promptContextChars` (4000), `promptMaxChars` (240), `rulesAutoActivate`
+(true), `rulesEnrichment` (true — mining on/off), `enrichmentEnabled` (true),
+`enrichmentIdleMs` / `enrichmentBatchCap` / `enrichmentMaxTokens`.
+
+## 7. Honest limits (stated, not buried)
+
+- **Run submits as-is.** No current host lets a docked plugin write the composer draft, so
+  click-to-edit a suggestion before running it is not available — the bar explains this where you'd
+  expect editing, and Run submits (#126 tracks the host ask). To change a suggestion: turn autorun
+  off and type your own version.
+- **Autorun never fires a `GOAL:` offer** — objective changes are human turns by host design.
+- **Nothing arrives armed.** Autorun is per-session, process-local; restarts and continuations
+  start disarmed. The goal is the durable half of the pair.
+- **A blocked or completed goal renders no strip** in the goal bar — if the bar vanished, the goal
+  reached a terminal state; `/chapters-prompt state` reports the whole picture (suggestion, model,
+  autorun, intent, goal) in one block.
+- **Displacement needs a host seam.** Where `GoalService.disarm` is absent, arming autorun
+  degrades safely: the loop yields to the goal driver instead of racing it, and the command's
+  reply tells you which behavior you got.
+- **The loop is as good as the prompt model.** Under autorun that model is the director of every
+  next step — the single most consequential of the three settings. `/chapters-models <provider>
+  <model>` shows its context window before you commit.

package/lib/prompt-state.js

@@ -59,7 +59,12 @@ export function goalDirectiveOf(goal) {
     if (goal === null)
         return 'No goal is active for this session. Recommend the next step the work itself calls for.';
     return [
-        `An ACTIVE GOAL governs this session (revision ${String(goal.revision)}): ${goal.objective.slice(0, 600)}`,
+        // #188 invariant: the objective rides WHOLE. The slice(0, 600) that stood here was itself the
+        // truncation the contract forbids — real objectives exceed 600 chars (the one this repo's own
+        // session ran under was ~1,200), and a director steering against half an objective is the exact
+        // hallucination invite zero-truncation exists to forbid. The assembly's budget handles size:
+        // the goal block is protected, and a protected set that cannot fit refuses with the numbers.
+        `An ACTIVE GOAL governs this session (revision ${String(goal.revision)}): ${goal.objective}`,
         'The recommended next step MUST advance that goal — a step that merely continues current work does not count.',
         'If the goal is already accomplished, recommend the step that closes it out (mark it complete) rather than another round.',
     ].join('\n');

package/lib/prompt-state.js.map

@@ -1 +1 @@
-{"version":3,"file":"prompt-state.js","sourceRoot":"","sources":["../src/prompt-state.ts"],"names":[],"mappings":"AAyBA,MAAM,KAAK,GAAG,CAAC,CAAU,EAAe,EAAE,CAAC,OAAO,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,CAAC,CAAA;AAEhF;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,UAAU,cAAc,CAAC,MAAmC;IAChE,IAAI,OAAO,GAAsB,IAAI,CAAA;IACrC,KAAK,MAAM,EAAE,IAAI,MAAM,EAAE,CAAC;QACxB,IAAI,EAAE,CAAC,IAAI,KAAK,aAAa;YAAE,SAAQ;QACvC,MAAM,IAAI,GAAG,CAAC,EAAE,CAAC,IAAI,IAAI,EAAE,CAA4B,CAAA;QACvD,MAAM,SAAS,GAAG,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,CAAA;QAC7D,IAAI,SAAS,KAAK,OAAO,IAAI,SAAS,KAAK,QAAQ,IAAI,SAAS,KAAK,QAAQ,EAAE,CAAC;YAAC,OAAO,GAAG,IAAI,CAAC;YAAC,SAAQ;QAAC,CAAC;QAC3G,MAAM,IAAI,GAAG,CAAC,IAAI,CAAC,IAAI,IAAI,EAAE,CAA4B,CAAA;QACzD,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC;YAAE,SAAQ;QACvD,MAAM,QAAQ,GAAG,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,QAAQ,CAAA;QAC1D,MAAM,KAAK,GAAkB,QAAQ,KAAK,QAAQ,CAAC,CAAC,CAAC,QAAQ;YAC3D,CAAC,CAAC,QAAQ,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS;gBACpC,CAAC,CAAC,QAAQ,KAAK,UAAU,CAAC,CAAC,CAAC,UAAU;oBACtC,CAAC,CAAC,QAAQ,CAAA;QACZ,MAAM,MAAM,GAAe;YACzB,EAAE,EAAE,IAAI,CAAC,EAAE;YACX,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,QAAkB,CAAC,CAAC,CAAC,CAAE,IAAI,CAAC,QAAmB,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,QAAQ,IAAI,CAAC,CAAC;YACzG,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,iGAAiG;YACjG,KAAK,EAAE,OAAO,KAAK,IAAI,IAAI,OAAO,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,IAAI,OAAO,CAAC,KAAK,KAAK,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,KAAK;YACtG,aAAa,EAAE,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,aAAuB,CAAC,CAAC,CAAC,CAAE,IAAI,CAAC,aAAwB,CAAC,CAAC,CAAC,IAAI;SACrG,CAAA;QACD,OAAO,GAAG,SAAS,KAAK,UAAU,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC,CAAC,MAAM,CAAA;IAChF,CAAC;IACD,OAAO,OAAO,KAAK,IAAI,IAAI,OAAO,CAAC,KAAK,KAAK,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,OAAO,CAAA;AAC1E,CAAC;AAED,iHAAiH;AACjH,MAAM,UAAU,qBAAqB,CAAC,IAAuB;IAC3D,OAAO,IAAI,KAAK,IAAI,IAAI,IAAI,CAAC,KAAK,KAAK,QAAQ,CAAA;AACjD,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,eAAe,CAAC,IAAuB;IACrD,IAAI,IAAI,KAAK,IAAI;QAAE,OAAO,wFAAwF,CAAA;IAClH,OAAO;QACL,iDAAiD,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE;QAC1G,8GAA8G;QAC9G,0HAA0H;KAC3H,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;AACd,CAAC;AA0BD,MAAM,CAAC,MAAM,YAAY,GACvB,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,MAAM,EAAE,aAAa,EAAE,YAAY,EAAE,IAAI,EAAE,CAAA;AAU1G;;;;;;;GAOG;AACH,MAAM,UAAU,aAAa,CAAC,GAAW;IACvC,MAAM,CAAC,GAAG,GAAG,CAAC,IAAI,EAAE,CAAA;IACpB,MAAM,CAAC,GAAG,eAAe,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;IACjC,IAAI,CAAC,KAAK,IAAI;QAAE,OAAO,EAAE,MAAM,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,EAAE,CAAA;IACzD,MAAM,IAAI,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,CAAA;IACxC,sGAAsG;IACtG,oDAAoD;IACpD,IAAI,IAAI,KAAK,EAAE;QAAE,OAAO,EAAE,MAAM,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,EAAE,CAAA;IAC1D,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE,CAAA;AAC5C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,qBAAqB,CAAC,IAAuB;IAC3D,OAAO,IAAI,KAAK,IAAI,IAAI,IAAI,CAAC,KAAK,KAAK,QAAQ,IAAI,IAAI,CAAC,UAAU,KAAK,OAAO,CAAA;AAChF,CAAC"}
+{"version":3,"file":"prompt-state.js","sourceRoot":"","sources":["../src/prompt-state.ts"],"names":[],"mappings":"AAyBA,MAAM,KAAK,GAAG,CAAC,CAAU,EAAe,EAAE,CAAC,OAAO,CAAC,KAAK,QAAQ,IAAI,CAAC,CAAC,MAAM,GAAG,CAAC,CAAA;AAEhF;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,UAAU,cAAc,CAAC,MAAmC;IAChE,IAAI,OAAO,GAAsB,IAAI,CAAA;IACrC,KAAK,MAAM,EAAE,IAAI,MAAM,EAAE,CAAC;QACxB,IAAI,EAAE,CAAC,IAAI,KAAK,aAAa;YAAE,SAAQ;QACvC,MAAM,IAAI,GAAG,CAAC,EAAE,CAAC,IAAI,IAAI,EAAE,CAA4B,CAAA;QACvD,MAAM,SAAS,GAAG,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,CAAA;QAC7D,IAAI,SAAS,KAAK,OAAO,IAAI,SAAS,KAAK,QAAQ,IAAI,SAAS,KAAK,QAAQ,EAAE,CAAC;YAAC,OAAO,GAAG,IAAI,CAAC;YAAC,SAAQ;QAAC,CAAC;QAC3G,MAAM,IAAI,GAAG,CAAC,IAAI,CAAC,IAAI,IAAI,EAAE,CAA4B,CAAA;QACzD,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC;YAAE,SAAQ;QACvD,MAAM,QAAQ,GAAG,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,QAAQ,CAAA;QAC1D,MAAM,KAAK,GAAkB,QAAQ,KAAK,QAAQ,CAAC,CAAC,CAAC,QAAQ;YAC3D,CAAC,CAAC,QAAQ,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS;gBACpC,CAAC,CAAC,QAAQ,KAAK,UAAU,CAAC,CAAC,CAAC,UAAU;oBACtC,CAAC,CAAC,QAAQ,CAAA;QACZ,MAAM,MAAM,GAAe;YACzB,EAAE,EAAE,IAAI,CAAC,EAAE;YACX,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,QAAkB,CAAC,CAAC,CAAC,CAAE,IAAI,CAAC,QAAmB,CAAC,CAAC,CAAC,CAAC,OAAO,EAAE,QAAQ,IAAI,CAAC,CAAC;YACzG,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,iGAAiG;YACjG,KAAK,EAAE,OAAO,KAAK,IAAI,IAAI,OAAO,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,IAAI,OAAO,CAAC,KAAK,KAAK,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,KAAK;YACtG,aAAa,EAAE,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,aAAuB,CAAC,CAAC,CAAC,CAAE,IAAI,CAAC,aAAwB,CAAC,CAAC,CAAC,IAAI;SACrG,CAAA;QACD,OAAO,GAAG,SAAS,KAAK,UAAU,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC,CAAC,CAAC,MAAM,CAAA;IAChF,CAAC;IACD,OAAO,OAAO,KAAK,IAAI,IAAI,OAAO,CAAC,KAAK,KAAK,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,OAAO,CAAA;AAC1E,CAAC;AAED,iHAAiH;AACjH,MAAM,UAAU,qBAAqB,CAAC,IAAuB;IAC3D,OAAO,IAAI,KAAK,IAAI,IAAI,IAAI,CAAC,KAAK,KAAK,QAAQ,CAAA;AACjD,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,eAAe,CAAC,IAAuB;IACrD,IAAI,IAAI,KAAK,IAAI;QAAE,OAAO,wFAAwF,CAAA;IAClH,OAAO;QACL,8FAA8F;QAC9F,8FAA8F;QAC9F,gGAAgG;QAChG,6FAA6F;QAC7F,6FAA6F;QAC7F,iDAAiD,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,IAAI,CAAC,SAAS,EAAE;QAC5F,8GAA8G;QAC9G,0HAA0H;KAC3H,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;AACd,CAAC;AA0BD,MAAM,CAAC,MAAM,YAAY,GACvB,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,EAAE,MAAM,EAAE,aAAa,EAAE,YAAY,EAAE,IAAI,EAAE,CAAA;AAU1G;;;;;;;GAOG;AACH,MAAM,UAAU,aAAa,CAAC,GAAW;IACvC,MAAM,CAAC,GAAG,GAAG,CAAC,IAAI,EAAE,CAAA;IACpB,MAAM,CAAC,GAAG,eAAe,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;IACjC,IAAI,CAAC,KAAK,IAAI;QAAE,OAAO,EAAE,MAAM,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,EAAE,CAAA;IACzD,MAAM,IAAI,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,CAAA;IACxC,sGAAsG;IACtG,oDAAoD;IACpD,IAAI,IAAI,KAAK,EAAE;QAAE,OAAO,EAAE,MAAM,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,EAAE,CAAA;IAC1D,OAAO,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,IAAI,EAAE,CAAA;AAC5C,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,qBAAqB,CAAC,IAAuB;IAC3D,OAAO,IAAI,KAAK,IAAI,IAAI,IAAI,CAAC,KAAK,KAAK,QAAQ,IAAI,IAAI,CAAC,UAAU,KAAK,OAAO,CAAA;AAChF,CAAC"}

package/lib/version.d.ts

@@ -6,4 +6,4 @@
  * reader cannot recover later, so it belongs in the bytes. A row in tests asserts this literal equals
  * `package.json`'s version — the one place a hand-mirrored constant must never be allowed to drift.
  */
-export declare const PLUGIN_VERSION = "0.2.5";
+export declare const PLUGIN_VERSION = "0.2.6";

package/lib/version.js

@@ -6,5 +6,5 @@
  * reader cannot recover later, so it belongs in the bytes. A row in tests asserts this literal equals
  * `package.json`'s version — the one place a hand-mirrored constant must never be allowed to drift.
  */
-export const PLUGIN_VERSION = '0.2.5';
+export const PLUGIN_VERSION = '0.2.6';
 //# sourceMappingURL=version.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@treeseed/dsh-chapters",
-  "version": "0.2.5",
+  "version": "0.2.6",
   "description": "dsh plugin: verbatim chapter archive + zero-token index for small-context models",
   "type": "module",
   "license": "Apache-2.0",